FoxPlan Agentic AI
FoxPlan integrates an agentic AI: an assistant that understands your requests in natural language and directly performs actions in your projects (create tasks, set dependencies, assign resources, shift a schedule…). Unlike a passive assistant that merely answers, the AI acts inside the application and then confirms what it did.
Configuration happens in Settings → Foxplan account → AI engine configuration.
This screen is reserved for Foxplan accounts (editor administration). By default, no AI engine is active: you must add one and enable it to allow AI requests in the application.
What the AI can do
The assistant only accesses predefined, secured business actions on the projects you have access to:
- Structure a project — generate a task hierarchy, dependencies and milestones from a simple description
- Assign resources — check availability then assign the right people to tasks
- Bulk-edit — for example "shift all tasks by two months", recomputing dates while preserving dependencies
- Manage tasks — list, create, edit, delete tasks and their due dates
- Comment — add comments on project items
The AI always works within your permission scope: it cannot do anything you couldn't do yourself. The AI authorizations (next tab) can narrow that scope further, company by company.
The configuration screen
Tailoring the AI to your company
Without configuration, the assistant answers like a generic tool: it knows FoxPlan, not your company. Three settings, which combine, make it specific to your organization.
| Setting | What it is | Where | When to use it |
|---|---|---|---|
| Organization context | A few lines of in-house vocabulary and internal rules, known to the AI in every answer | Context tab | Conventions, thresholds, the way you name things |
| Internal knowledge base | Full documents the AI consults on demand | Knowledge tab | Internal rules, methodology, quality charter, procedures |
| Predefined tasks | Ready-to-run requests, available from the chat | Predefined tasks tab | Recurring deliverables: status report, steering committee deck, risk review |
On top of that, each user writes their own personal instructions in their profile.
A short rule that must always apply belongs in the organization context. A ten-page document belongs in the knowledge base: the AI will only read the relevant passages, instead of re-reading everything on every question. A recurring request becomes a predefined task.
The screen has seven tabs.
AI engine configuration (Parameters)
This is where you declare the LLM used by your company. The table lists the configured engines (provider, name, model, presence of an API key, status). The Add an AI engine button opens the following form:
| Field | Description |
|---|---|
| Provider | OpenAI, Anthropic, Azure OpenAI, Microsoft Copilot or Custom |
| Display name | Label identifying the configuration in the list |
| API key | Authentication with the provider. Encrypted at rest, never returned to the browser after saving (leave empty to keep the existing key) |
| Base URL | API endpoint (useful for a Custom or self-hosted endpoint) |
| Model | Exact model identifier at the provider (e.g. gpt-4o, claude-sonnet-4-6, gpt-4o-mini) |
| Temperature | Randomness of responses (0 to 2). Low = more factual, high = more creative. 0.2 recommended |
| Max tokens | Maximum length of a response (1 token ≈ ¾ of a word). Caps costs |
| Azure deployment name / API version | Specific to the Azure OpenAI provider |
| Additional HTTP headers | For a Custom provider (one per line, format Name=value) |
| Active configuration | Enables this engine. Only one configuration can be active at a time — enabling one disables the others |
Each model has its own capabilities, context window, token limits and pricing. Refer to the provider's documentation for the up-to-date list of model identifiers.
Organization context
This is where the AI learns who you are. Two free-text fields, filled in by the administrator and passed to the assistant on every question asked by a user of the company.
| Field | What goes in it | Example |
|---|---|---|
| Identity and vocabulary | What the company does, how it names things, its units and conventions | "We are a civil engineering design office. We call 'site' what FoxPlan calls a project. Our phases are APS, APD, PRO, EXE. Amounts are in euros excluding tax." |
| Internal rules and governance | Thresholds, approval paths and steering conventions the assistant must apply | "Any risk with a criticality of 12 or above is escalated to the steering committee. A project status turns red as soon as a contractual milestone slips." |
Each field is limited to 3,000 characters, and a counter tells you where you stand. This limit is not arbitrary: the text travels with every request sent to the model, so it is billed every time. Stick to short, general rules — detailed documents belong in the knowledge base, in the Knowledge tab.
This text changes how the AI answers. It grants no additional data access and lifts no restriction set in the AI authorizations tab.
Internal knowledge base
In the Knowledge tab, you upload the documents the AI may consult: internal rules, project methodology, quality charter, procurement reference, procedures.
How it works: on upload, FoxPlan extracts the text from the file (Word, PDF, Excel, PowerPoint, text, OpenDocument) and keeps only that — the original file is not stored. The assistant permanently sees only the title and purpose of each document; it looks up the relevant passages only when the question calls for it. That is what makes large corpora possible without weighing down every exchange.
| Item | Detail |
|---|---|
| Purpose of the document | The most important field on the screen: the AI decides whether to open a document based on it. Write "Leave, expense reports and remote work" rather than "HR document" |
| Maximum size | 10 MB per file; 200,000 characters of text kept |
| Number of documents | 50 per company |
| Enable / disable | A disabled document disappears from the AI's searches without being deleted |
The AI picks the documents to open from their purpose. Two minutes spent writing a good description beat a perfect document the assistant never opens.
Uploaded content is treated as reference material, never as instructions: a directive slipped into a document changes neither the assistant's behaviour nor the user's permissions.
Predefined tasks
A predefined task (or skill) is a request written once by the administrator, that users then run in one click from the AI chat.
Each predefined task has:
| Field | Role |
|---|---|
| Name | The label shown in the chat drop-down (e.g. Project status report) |
| What it does | A help sentence displayed under the name |
| Instructions given to the AI | The core: the outline, tone, format and data expected. Up to 5,000 characters |
| Offer to users | An inactive predefined task disappears from the list without being deleted |
Example instructions:
Write the project status report in three parts — progress (with completion and this month's milestones), open risks sorted by criticality, decisions expected from the steering committee. End with an overall colour (green/amber/red) justified in one sentence.
On the user side, the Predefined task drop-down appears below the input area of the AI chat (it stays hidden until at least one task is published). The user picks one, optionally adds a detail — "on the IS Overhaul project, for June" — and sends. The predefined task applies to that message only: it clears itself afterwards, so it never silently shapes the next question.
It states what is expected from the AI, nothing more. The user's permissions, the organization context and the authorizations keep applying exactly the same way.
AI authorizations
This tab defines what the AI is allowed to do in the application. Operations are grouped into categories, and each category is authorized separately for the two channels through which the AI can act:
| Column | What it covers |
|---|---|
| AI engine | The assistant built into FoxPlan |
| MCP server | External MCP clients connected to FoxPlan's MCP server (Claude, for example) |
| Category | Operations covered |
|---|---|
| Read data | Read projects, tasks, members, resources, risks, decisions and other project objects |
| Read documentation | Search and read the FoxPlan user documentation |
| Internal company knowledge | Read the organization context and the documents of the internal knowledge base |
| Create tasks | Create tasks, generate them from a WBS, copy them across projects, create dependency links |
| Update and delete tasks | Update or delete tasks, change deadlines, remove allocations |
| Manage projects | Create and update projects, add members to a project |
| Assign resources | Assign a resource to a task |
| Create cards and tickets | Create cards on the personal kanban or on a project |
| Add comments | Add comments on tasks |
| Export documents | Generate Word, Excel, PowerPoint or PDF documents |
The checkbox in a column header allows or denies the whole channel in one go. Changes only take effect once you click Save.
As long as no restriction has been saved, every operation is permitted on both channels: existing companies keep their current behaviour. This grid is there to restrict, never to open up.
Ticking a box grants no extra right: the operation remains subject to the user's permissions on the project concerned. Ticking a category merely lifts a ban set at company level.
When a request targets an unticked operation, it is not executed: the AI explains this to the user in their language and does not try to work around the restriction another way. The check applies however the request is phrased.
Only company administrators can change this grid.
AI request history
The detailed log of every call to the AI engine made by the company's users. For each request: date, user, provider, model, project, tokens (prompt / completion / total), duration, status and any error message, language.
Retention: 24 months. Useful to audit usage and diagnose a failed request.
The log also keeps the predefined task that was run, when there is one: since the instructions sent to the model vary from one message to the next, this is often what explains an unexpected answer.
Consumption summary
The consolidated monthly consumption of AI tokens, across all users and models: per month, number of requests (successful / failed) and tokens (prompt / completion / total).
This is the view to track costs and anticipate billing, since AI usage is billed per token by the providers.
What the context costs in tokens
AI engines bill per token, and everything you configure here ends up, in one form or another, in what gets sent to the model. Here is where the money goes (1 token ≈ ¾ of a word ≈ 4 characters; the figures are orders of magnitude).
| Item | What it adds | When |
|---|---|---|
| Organization context | ~200 tokens for 600 to 800 characters; ~1,500 tokens if both fields are filled to the maximum | On every request, for every user |
| Personal instructions | up to ~500 tokens | On every request from that user |
| Knowledge base summary | ~30 tokens per document (title + purpose): ~300 tokens for 10 documents, ~1,500 at the cap of 50 | On every request |
| Document excerpts | ~1,200 tokens (up to 4 excerpts); ~2,000 tokens if the AI reads a full page | Only when the question triggers a search |
| Predefined task | up to ~1,250 tokens | Only when it is run |
To put those numbers in perspective: the technical instructions FoxPlan already sends on every request (how to call the tools, how to format the answer) amount on their own to roughly 1,400 tokens, on top of which come the descriptions of the available tools. A properly configured company adds about 500 tokens per request — a 10 to 15 % increase in input cost: real, but marginal.
On a request where the AI acts, the full prompt is sent back to the model at every tool round-trip. A 500-token context, in a conversation that chains five tool calls, costs 2,500 tokens rather than 500. That is precisely why document content is never injected into the prompt: at that rate, a thirty-page set of internal rules would cost more than the answer itself.
Three habits that keep the bill down
- Short rules in the context, details in the documents. The context is paid for on every question from every user; a document is paid for only when it is used.
- Do not upload fifty documents "just in case". Every active document adds a line to the summary, hence a permanent cost, even if it is never opened. Disable the ones that no longer serve.
- Keep the purpose short and precise. It is paid for on every request and it is what decides whether the document gets opened: two well-phrased lines beat a paragraph.
The Consumption summary tab gives you the real measurement on your own usage: compare a month before and a month after configuring the context — that is the only figure that truly matters.
Security and privacy
- API keys are encrypted at rest and never re-exposed to the browser.
- The AI only executes validated business actions, within the connected user's permissions.
- Every call is traced in the history for audit.
- The AI authorizations let you disable, per company and per channel, the categories of operations the AI must not perform.
- The context, the documents and the predefined tasks are partitioned by company: they are only visible to users of the same workspace.
- The content of uploaded documents is treated as reference material, never as instructions: a directive slipped into it changes neither the assistant's behaviour nor the user's permissions.
- A user's personal instructions apply to their own conversations only, and yield to the organization context in case of contradiction.
What about external MCP clients?
An MCP client (Claude, for example) connected to FoxPlan's MCP server does not receive the internal instructions of the FoxPlan assistant. It does have a dedicated tool, getCompanyContext, which returns the company's vocabulary and rules along with the connected user's personal instructions, as well as knowledge base search tools (searchCompanyKnowledge, getCompanyKnowledgeDocument).
These tools belong to the Internal company knowledge category: unchecking that category in the MCP server column closes access to external clients while leaving it open to the FoxPlan assistant.
Predefined tasks, on the other hand, are specific to the FoxPlan chat and are not exposed to MCP clients.
See also
- Foxplan account — the parent screen hosting the AI configuration
- My instructions to the AI assistant — personalization on the user side
- API usage guide — for programmatic integrations
To enable the feature on your account, contact the FoxPlan team.